Exploration of multivariate atmospheric particulate compositional data by projection pursuit

Abstract

With the availability of analytical methods that can measure multiple chemical species in a single sample and automated samplers that can obtain large numbers of samples, it is possible to obtain large data sets characterizing some portion of the atmospheric environment. The problem that arises with these large multivariate data sets is that they are too big to be easily visualized and numerical methods are needed to permit the identification of structures that exist in such data. One of these exploratory analysis methods is called projection pursuit. Projection pursuit is a method for projecting high dimension data into fewer and hence comprehensible number of dimensions that capture the main features of the data. Projection pursuit is a relatively new method for identifying interesting lower-dimensional views of multivariate data by optimizing a criterion index that measures how interesting each projection is. The basic methodology is outlined and its application to coarse and fine particle composition data obtained during the Regional Air Pollution Study of St. Louis, Missouri, are presented. The method permits the identification of clusters of samples that are influenced by the same emission sources. By examining the average wind direction for the clusters, it is possible to identify the individual sources that contribute material to those samples.